diff --git a/vendor/github.com/klauspost/compress/huff0/.gitignore b/vendor/github.com/klauspost/compress/huff0/.gitignore
new file mode 100644
index 0000000..b3d2629
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/.gitignore
@@ -0,0 +1 @@
+/huff0-fuzz.zip
diff --git a/vendor/github.com/klauspost/compress/huff0/README.md b/vendor/github.com/klauspost/compress/huff0/README.md
new file mode 100644
index 0000000..0a8448c
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/README.md
@@ -0,0 +1,87 @@
+# Huff0 entropy compression
+
+This package provides Huff0 encoding and decoding as used in zstd.
+            
+[Huff0](https://github.com/Cyan4973/FiniteStateEntropy#new-generation-entropy-coders), 
+a Huffman codec designed for modern CPU, featuring OoO (Out of Order) operations on multiple ALU 
+(Arithmetic Logic Unit), achieving extremely fast compression and decompression speeds.
+
+This can be used for compressing input with a lot of similar input values to the smallest number of bytes.
+This does not perform any multi-byte [dictionary coding](https://en.wikipedia.org/wiki/Dictionary_coder) as LZ coders,
+but it can be used as a secondary step to compressors (like Snappy) that does not do entropy encoding. 
+
+* [Godoc documentation](https://godoc.org/github.com/klauspost/compress/huff0)
+
+THIS PACKAGE IS NOT CONSIDERED STABLE AND API OR ENCODING MAY CHANGE IN THE FUTURE.
+
+## News
+
+ * Mar 2018: First implementation released. Consider this beta software for now.
+
+# Usage
+
+This package provides a low level interface that allows to compress single independent blocks. 
+
+Each block is separate, and there is no built in integrity checks. 
+This means that the caller should keep track of block sizes and also do checksums if needed.  
+
+Compressing a block is done via the [`Compress1X`](https://godoc.org/github.com/klauspost/compress/huff0#Compress1X) and 
+[`Compress4X`](https://godoc.org/github.com/klauspost/compress/huff0#Compress4X) functions.
+You must provide input and will receive the output and maybe an error.
+
+These error values can be returned:
+
+| Error               | Description                                                                 |
+|---------------------|-----------------------------------------------------------------------------|
+| `<nil>`             | Everything ok, output is returned                                           |
+| `ErrIncompressible` | Returned when input is judged to be too hard to compress                    |
+| `ErrUseRLE`         | Returned from the compressor when the input is a single byte value repeated |
+| `ErrTooBig`         | Returned if the input block exceeds the maximum allowed size (128 Kib)      |
+| `(error)`           | An internal error occurred.                                                 |
+
+
+As can be seen above some of there are errors that will be returned even under normal operation so it is important to handle these.
+
+To reduce allocations you can provide a [`Scratch`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch) object 
+that can be re-used for successive calls. Both compression and decompression accepts a `Scratch` object, and the same 
+object can be used for both.   
+
+Be aware, that when re-using a `Scratch` object that the *output* buffer is also re-used, so if you are still using this
+you must set the `Out` field in the scratch to nil. The same buffer is used for compression and decompression output.
+
+The `Scratch` object will retain state that allows to re-use previous tables for encoding and decoding.  
+
+## Tables and re-use
+
+Huff0 allows for reusing tables from the previous block to save space if that is expected to give better/faster results. 
+
+The Scratch object allows you to set a [`ReusePolicy`](https://godoc.org/github.com/klauspost/compress/huff0#ReusePolicy) 
+that controls this behaviour. See the documentation for details. This can be altered between each block.
+
+Do however note that this information is *not* stored in the output block and it is up to the users of the package to
+record whether [`ReadTable`](https://godoc.org/github.com/klauspost/compress/huff0#ReadTable) should be called,
+based on the boolean reported back from the CompressXX call. 
+
+If you want to store the table separate from the data, you can access them as `OutData` and `OutTable` on the 
+[`Scratch`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch) object.
+
+## Decompressing
+
+The first part of decoding is to initialize the decoding table through [`ReadTable`](https://godoc.org/github.com/klauspost/compress/huff0#ReadTable).
+This will initialize the decoding tables. 
+You can supply the complete block to `ReadTable` and it will return the data part of the block 
+which can be given to the decompressor. 
+
+Decompressing is done by calling the [`Decompress1X`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch.Decompress1X) 
+or [`Decompress4X`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch.Decompress4X) function.
+
+You must provide the output from the compression stage, at exactly the size you got back. If you receive an error back
+your input was likely corrupted. 
+
+It is important to note that a successful decoding does *not* mean your output matches your original input. 
+There are no integrity checks, so relying on errors from the decompressor does not assure your data is valid.
+
+# Contributing
+
+Contributions are always welcome. Be aware that adding public functions will require good justification and breaking 
+changes will likely not be accepted. If in doubt open an issue before writing the PR.
\ No newline at end of file
diff --git a/vendor/github.com/klauspost/compress/huff0/bitreader.go b/vendor/github.com/klauspost/compress/huff0/bitreader.go
new file mode 100644
index 0000000..7d0903c
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/bitreader.go
@@ -0,0 +1,115 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package huff0
+
+import (
+	"errors"
+	"io"
+)
+
+// bitReader reads a bitstream in reverse.
+// The last set bit indicates the start of the stream and is used
+// for aligning the input.
+type bitReader struct {
+	in       []byte
+	off      uint // next byte to read is at in[off - 1]
+	value    uint64
+	bitsRead uint8
+}
+
+// init initializes and resets the bit reader.
+func (b *bitReader) init(in []byte) error {
+	if len(in) < 1 {
+		return errors.New("corrupt stream: too short")
+	}
+	b.in = in
+	b.off = uint(len(in))
+	// The highest bit of the last byte indicates where to start
+	v := in[len(in)-1]
+	if v == 0 {
+		return errors.New("corrupt stream, did not find end of stream")
+	}
+	b.bitsRead = 64
+	b.value = 0
+	b.fill()
+	b.fill()
+	b.bitsRead += 8 - uint8(highBit32(uint32(v)))
+	return nil
+}
+
+// getBits will return n bits. n can be 0.
+func (b *bitReader) getBits(n uint8) uint16 {
+	if n == 0 || b.bitsRead >= 64 {
+		return 0
+	}
+	return b.getBitsFast(n)
+}
+
+// getBitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReader) getBitsFast(n uint8) uint16 {
+	const regMask = 64 - 1
+	v := uint16((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask))
+	b.bitsRead += n
+	return v
+}
+
+// peekBitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReader) peekBitsFast(n uint8) uint16 {
+	const regMask = 64 - 1
+	v := uint16((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask))
+	return v
+}
+
+// fillFast() will make sure at least 32 bits are available.
+// There must be at least 4 bytes available.
+func (b *bitReader) fillFast() {
+	if b.bitsRead < 32 {
+		return
+	}
+	// Do single re-slice to avoid bounds checks.
+	v := b.in[b.off-4 : b.off]
+	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	b.value = (b.value << 32) | uint64(low)
+	b.bitsRead -= 32
+	b.off -= 4
+}
+
+// fill() will make sure at least 32 bits are available.
+func (b *bitReader) fill() {
+	if b.bitsRead < 32 {
+		return
+	}
+	if b.off > 4 {
+		v := b.in[b.off-4 : b.off]
+		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+		b.value = (b.value << 32) | uint64(low)
+		b.bitsRead -= 32
+		b.off -= 4
+		return
+	}
+	for b.off > 0 {
+		b.value = (b.value << 8) | uint64(b.in[b.off-1])
+		b.bitsRead -= 8
+		b.off--
+	}
+}
+
+// finished returns true if all bits have been read from the bit stream.
+func (b *bitReader) finished() bool {
+	return b.off == 0 && b.bitsRead >= 64
+}
+
+// close the bitstream and returns an error if out-of-buffer reads occurred.
+func (b *bitReader) close() error {
+	// Release reference.
+	b.in = nil
+	if b.bitsRead > 64 {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/bitwriter.go b/vendor/github.com/klauspost/compress/huff0/bitwriter.go
new file mode 100644
index 0000000..ec0c3fc
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/bitwriter.go
@@ -0,0 +1,186 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package huff0
+
+import "fmt"
+
+// bitWriter will write bits.
+// First bit will be LSB of the first byte of output.
+type bitWriter struct {
+	bitContainer uint64
+	nBits        uint8
+	out          []byte
+}
+
+// bitMask16 is bitmasks. Has extra to avoid bounds check.
+var bitMask16 = [32]uint16{
+	0, 1, 3, 7, 0xF, 0x1F,
+	0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF,
+	0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF, 0xFFFF,
+	0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
+	0xFFFF, 0xFFFF} /* up to 16 bits */
+
+// addBits16NC will add up to 16 bits.
+// It will not check if there is space for them,
+// so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits16NC(value uint16, bits uint8) {
+	b.bitContainer |= uint64(value&bitMask16[bits&31]) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated.
+// It will not check if there is space for them, so the caller must ensure that it has flushed recently.
+func (b *bitWriter) addBits16Clean(value uint16, bits uint8) {
+	b.bitContainer |= uint64(value) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated.
+// It will not check if there is space for them, so the caller must ensure that it has flushed recently.
+func (b *bitWriter) encSymbol(ct cTable, symbol byte) {
+	enc := ct[symbol]
+	b.bitContainer |= uint64(enc.val) << (b.nBits & 63)
+	b.nBits += enc.nBits
+}
+
+// addBits16ZeroNC will add up to 16 bits.
+// It will not check if there is space for them,
+// so the caller must ensure that it has flushed recently.
+// This is fastest if bits can be zero.
+func (b *bitWriter) addBits16ZeroNC(value uint16, bits uint8) {
+	if bits == 0 {
+		return
+	}
+	value <<= (16 - bits) & 15
+	value >>= (16 - bits) & 15
+	b.bitContainer |= uint64(value) << (b.nBits & 63)
+	b.nBits += bits
+}
+
+// flush will flush all pending full bytes.
+// There will be at least 56 bits available for writing when this has been called.
+// Using flush32 is faster, but leaves less space for writing.
+func (b *bitWriter) flush() {
+	v := b.nBits >> 3
+	switch v {
+	case 0:
+		return
+	case 1:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+		)
+		b.bitContainer >>= 1 << 3
+	case 2:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+		)
+		b.bitContainer >>= 2 << 3
+	case 3:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+		)
+		b.bitContainer >>= 3 << 3
+	case 4:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+		)
+		b.bitContainer >>= 4 << 3
+	case 5:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+		)
+		b.bitContainer >>= 5 << 3
+	case 6:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+		)
+		b.bitContainer >>= 6 << 3
+	case 7:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+			byte(b.bitContainer>>48),
+		)
+		b.bitContainer >>= 7 << 3
+	case 8:
+		b.out = append(b.out,
+			byte(b.bitContainer),
+			byte(b.bitContainer>>8),
+			byte(b.bitContainer>>16),
+			byte(b.bitContainer>>24),
+			byte(b.bitContainer>>32),
+			byte(b.bitContainer>>40),
+			byte(b.bitContainer>>48),
+			byte(b.bitContainer>>56),
+		)
+		b.bitContainer = 0
+		b.nBits = 0
+		return
+	default:
+		panic(fmt.Errorf("bits (%d) > 64", b.nBits))
+	}
+	b.nBits &= 7
+}
+
+// flush32 will flush out, so there are at least 32 bits available for writing.
+func (b *bitWriter) flush32() {
+	if b.nBits < 32 {
+		return
+	}
+	b.out = append(b.out,
+		byte(b.bitContainer),
+		byte(b.bitContainer>>8),
+		byte(b.bitContainer>>16),
+		byte(b.bitContainer>>24))
+	b.nBits -= 32
+	b.bitContainer >>= 32
+}
+
+// flushAlign will flush remaining full bytes and align to next byte boundary.
+func (b *bitWriter) flushAlign() {
+	nbBytes := (b.nBits + 7) >> 3
+	for i := uint8(0); i < nbBytes; i++ {
+		b.out = append(b.out, byte(b.bitContainer>>(i*8)))
+	}
+	b.nBits = 0
+	b.bitContainer = 0
+}
+
+// close will write the alignment bit and write the final byte(s)
+// to the output.
+func (b *bitWriter) close() error {
+	// End mark
+	b.addBits16Clean(1, 1)
+	// flush until next byte.
+	b.flushAlign()
+	return nil
+}
+
+// reset and continue writing by appending to out.
+func (b *bitWriter) reset(out []byte) {
+	b.bitContainer = 0
+	b.nBits = 0
+	b.out = out
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/bytereader.go b/vendor/github.com/klauspost/compress/huff0/bytereader.go
new file mode 100644
index 0000000..50bcdf6
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/bytereader.go
@@ -0,0 +1,54 @@
+// Copyright 2018 Klaus Post. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
+
+package huff0
+
+// byteReader provides a byte reader that reads
+// little endian values from a byte stream.
+// The input stream is manually advanced.
+// The reader performs no bounds checks.
+type byteReader struct {
+	b   []byte
+	off int
+}
+
+// init will initialize the reader and set the input.
+func (b *byteReader) init(in []byte) {
+	b.b = in
+	b.off = 0
+}
+
+// advance the stream b n bytes.
+func (b *byteReader) advance(n uint) {
+	b.off += int(n)
+}
+
+// Int32 returns a little endian int32 starting at current offset.
+func (b byteReader) Int32() int32 {
+	v3 := int32(b.b[b.off+3])
+	v2 := int32(b.b[b.off+2])
+	v1 := int32(b.b[b.off+1])
+	v0 := int32(b.b[b.off])
+	return (v3 << 24) | (v2 << 16) | (v1 << 8) | v0
+}
+
+// Uint32 returns a little endian uint32 starting at current offset.
+func (b byteReader) Uint32() uint32 {
+	v3 := uint32(b.b[b.off+3])
+	v2 := uint32(b.b[b.off+2])
+	v1 := uint32(b.b[b.off+1])
+	v0 := uint32(b.b[b.off])
+	return (v3 << 24) | (v2 << 16) | (v1 << 8) | v0
+}
+
+// unread returns the unread portion of the input.
+func (b byteReader) unread() []byte {
+	return b.b[b.off:]
+}
+
+// remain will return the number of bytes remaining.
+func (b byteReader) remain() int {
+	return len(b.b) - b.off
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/compress.go b/vendor/github.com/klauspost/compress/huff0/compress.go
new file mode 100644
index 0000000..dd4f7fe
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/compress.go
@@ -0,0 +1,618 @@
+package huff0
+
+import (
+	"fmt"
+	"runtime"
+	"sync"
+)
+
+// Compress1X will compress the input.
+// The output can be decoded using Decompress1X.
+// Supply a Scratch object. The scratch object contains state about re-use,
+// So when sharing across independent encodes, be sure to set the re-use policy.
+func Compress1X(in []byte, s *Scratch) (out []byte, reUsed bool, err error) {
+	s, err = s.prepare(in)
+	if err != nil {
+		return nil, false, err
+	}
+	return compress(in, s, s.compress1X)
+}
+
+// Compress4X will compress the input. The input is split into 4 independent blocks
+// and compressed similar to Compress1X.
+// The output can be decoded using Decompress4X.
+// Supply a Scratch object. The scratch object contains state about re-use,
+// So when sharing across independent encodes, be sure to set the re-use policy.
+func Compress4X(in []byte, s *Scratch) (out []byte, reUsed bool, err error) {
+	s, err = s.prepare(in)
+	if err != nil {
+		return nil, false, err
+	}
+	if false {
+		// TODO: compress4Xp only slightly faster.
+		const parallelThreshold = 8 << 10
+		if len(in) < parallelThreshold || runtime.GOMAXPROCS(0) == 1 {
+			return compress(in, s, s.compress4X)
+		}
+		return compress(in, s, s.compress4Xp)
+	}
+	return compress(in, s, s.compress4X)
+}
+
+func compress(in []byte, s *Scratch, compressor func(src []byte) ([]byte, error)) (out []byte, reUsed bool, err error) {
+	// Nuke previous table if we cannot reuse anyway.
+	if s.Reuse == ReusePolicyNone {
+		s.prevTable = s.prevTable[:0]
+	}
+
+	// Create histogram, if none was provided.
+	maxCount := s.maxCount
+	var canReuse = false
+	if maxCount == 0 {
+		maxCount, canReuse = s.countSimple(in)
+	} else {
+		canReuse = s.canUseTable(s.prevTable)
+	}
+
+	// Reset for next run.
+	s.clearCount = true
+	s.maxCount = 0
+	if maxCount >= len(in) {
+		if maxCount > len(in) {
+			return nil, false, fmt.Errorf("maxCount (%d) > length (%d)", maxCount, len(in))
+		}
+		if len(in) == 1 {
+			return nil, false, ErrIncompressible
+		}
+		// One symbol, use RLE
+		return nil, false, ErrUseRLE
+	}
+	if maxCount == 1 || maxCount < (len(in)>>7) {
+		// Each symbol present maximum once or too well distributed.
+		return nil, false, ErrIncompressible
+	}
+
+	if s.Reuse == ReusePolicyPrefer && canReuse {
+		keepTable := s.cTable
+		s.cTable = s.prevTable
+		s.Out, err = compressor(in)
+		s.cTable = keepTable
+		if err == nil && len(s.Out) < len(in) {
+			s.OutData = s.Out
+			return s.Out, true, nil
+		}
+		// Do not attempt to re-use later.
+		s.prevTable = s.prevTable[:0]
+	}
+
+	// Calculate new table.
+	s.optimalTableLog()
+	err = s.buildCTable()
+	if err != nil {
+		return nil, false, err
+	}
+
+	if false && !s.canUseTable(s.cTable) {
+		panic("invalid table generated")
+	}
+
+	if s.Reuse == ReusePolicyAllow && canReuse {
+		hSize := len(s.Out)
+		oldSize := s.prevTable.estimateSize(s.count[:s.symbolLen])
+		newSize := s.cTable.estimateSize(s.count[:s.symbolLen])
+		if oldSize <= hSize+newSize || hSize+12 >= len(in) {
+			// Retain cTable even if we re-use.
+			keepTable := s.cTable
+			s.cTable = s.prevTable
+			s.Out, err = compressor(in)
+			s.cTable = keepTable
+			if len(s.Out) >= len(in) {
+				return nil, false, ErrIncompressible
+			}
+			s.OutData = s.Out
+			return s.Out, true, nil
+		}
+	}
+
+	// Use new table
+	err = s.cTable.write(s)
+	if err != nil {
+		s.OutTable = nil
+		return nil, false, err
+	}
+	s.OutTable = s.Out
+
+	// Compress using new table
+	s.Out, err = compressor(in)
+	if err != nil {
+		s.OutTable = nil
+		return nil, false, err
+	}
+	if len(s.Out) >= len(in) {
+		s.OutTable = nil
+		return nil, false, ErrIncompressible
+	}
+	// Move current table into previous.
+	s.prevTable, s.cTable = s.cTable, s.prevTable[:0]
+	s.OutData = s.Out[len(s.OutTable):]
+	return s.Out, false, nil
+}
+
+func (s *Scratch) compress1X(src []byte) ([]byte, error) {
+	return s.compress1xDo(s.Out, src)
+}
+
+func (s *Scratch) compress1xDo(dst, src []byte) ([]byte, error) {
+	var bw = bitWriter{out: dst}
+
+	// N is length divisible by 4.
+	n := len(src)
+	n -= n & 3
+	cTable := s.cTable[:256]
+
+	// Encode last bytes.
+	for i := len(src) & 3; i > 0; i-- {
+		bw.encSymbol(cTable, src[n+i-1])
+	}
+	if s.actualTableLog <= 8 {
+		n -= 4
+		for ; n >= 0; n -= 4 {
+			tmp := src[n : n+4]
+			// tmp should be len 4
+			bw.flush32()
+			bw.encSymbol(cTable, tmp[3])
+			bw.encSymbol(cTable, tmp[2])
+			bw.encSymbol(cTable, tmp[1])
+			bw.encSymbol(cTable, tmp[0])
+		}
+	} else {
+		n -= 4
+		for ; n >= 0; n -= 4 {
+			tmp := src[n : n+4]
+			// tmp should be len 4
+			bw.flush32()
+			bw.encSymbol(cTable, tmp[3])
+			bw.encSymbol(cTable, tmp[2])
+			bw.flush32()
+			bw.encSymbol(cTable, tmp[1])
+			bw.encSymbol(cTable, tmp[0])
+		}
+	}
+	err := bw.close()
+	return bw.out, err
+}
+
+var sixZeros [6]byte
+
+func (s *Scratch) compress4X(src []byte) ([]byte, error) {
+	if len(src) < 12 {
+		return nil, ErrIncompressible
+	}
+	segmentSize := (len(src) + 3) / 4
+
+	// Add placeholder for output length
+	offsetIdx := len(s.Out)
+	s.Out = append(s.Out, sixZeros[:]...)
+
+	for i := 0; i < 4; i++ {
+		toDo := src
+		if len(toDo) > segmentSize {
+			toDo = toDo[:segmentSize]
+		}
+		src = src[len(toDo):]
+
+		var err error
+		idx := len(s.Out)
+		s.Out, err = s.compress1xDo(s.Out, toDo)
+		if err != nil {
+			return nil, err
+		}
+		// Write compressed length as little endian before block.
+		if i < 3 {
+			// Last length is not written.
+			length := len(s.Out) - idx
+			s.Out[i*2+offsetIdx] = byte(length)
+			s.Out[i*2+offsetIdx+1] = byte(length >> 8)
+		}
+	}
+
+	return s.Out, nil
+}
+
+// compress4Xp will compress 4 streams using separate goroutines.
+func (s *Scratch) compress4Xp(src []byte) ([]byte, error) {
+	if len(src) < 12 {
+		return nil, ErrIncompressible
+	}
+	// Add placeholder for output length
+	s.Out = s.Out[:6]
+
+	segmentSize := (len(src) + 3) / 4
+	var wg sync.WaitGroup
+	var errs [4]error
+	wg.Add(4)
+	for i := 0; i < 4; i++ {
+		toDo := src
+		if len(toDo) > segmentSize {
+			toDo = toDo[:segmentSize]
+		}
+		src = src[len(toDo):]
+
+		// Separate goroutine for each block.
+		go func(i int) {
+			s.tmpOut[i], errs[i] = s.compress1xDo(s.tmpOut[i][:0], toDo)
+			wg.Done()
+		}(i)
+	}
+	wg.Wait()
+	for i := 0; i < 4; i++ {
+		if errs[i] != nil {
+			return nil, errs[i]
+		}
+		o := s.tmpOut[i]
+		// Write compressed length as little endian before block.
+		if i < 3 {
+			// Last length is not written.
+			s.Out[i*2] = byte(len(o))
+			s.Out[i*2+1] = byte(len(o) >> 8)
+		}
+
+		// Write output.
+		s.Out = append(s.Out, o...)
+	}
+	return s.Out, nil
+}
+
+// countSimple will create a simple histogram in s.count.
+// Returns the biggest count.
+// Does not update s.clearCount.
+func (s *Scratch) countSimple(in []byte) (max int, reuse bool) {
+	reuse = true
+	for _, v := range in {
+		s.count[v]++
+	}
+	m := uint32(0)
+	if len(s.prevTable) > 0 {
+		for i, v := range s.count[:] {
+			if v > m {
+				m = v
+			}
+			if v > 0 {
+				s.symbolLen = uint16(i) + 1
+				if i >= len(s.prevTable) {
+					reuse = false
+				} else {
+					if s.prevTable[i].nBits == 0 {
+						reuse = false
+					}
+				}
+			}
+		}
+		return int(m), reuse
+	}
+	for i, v := range s.count[:] {
+		if v > m {
+			m = v
+		}
+		if v > 0 {
+			s.symbolLen = uint16(i) + 1
+		}
+	}
+	return int(m), false
+}
+
+func (s *Scratch) canUseTable(c cTable) bool {
+	if len(c) < int(s.symbolLen) {
+		return false
+	}
+	for i, v := range s.count[:s.symbolLen] {
+		if v != 0 && c[i].nBits == 0 {
+			return false
+		}
+	}
+	return true
+}
+
+// minTableLog provides the minimum logSize to safely represent a distribution.
+func (s *Scratch) minTableLog() uint8 {
+	minBitsSrc := highBit32(uint32(s.br.remain()-1)) + 1
+	minBitsSymbols := highBit32(uint32(s.symbolLen-1)) + 2
+	if minBitsSrc < minBitsSymbols {
+		return uint8(minBitsSrc)
+	}
+	return uint8(minBitsSymbols)
+}
+
+// optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog
+func (s *Scratch) optimalTableLog() {
+	tableLog := s.TableLog
+	minBits := s.minTableLog()
+	maxBitsSrc := uint8(highBit32(uint32(s.br.remain()-1))) - 2
+	if maxBitsSrc < tableLog {
+		// Accuracy can be reduced
+		tableLog = maxBitsSrc
+	}
+	if minBits > tableLog {
+		tableLog = minBits
+	}
+	// Need a minimum to safely represent all symbol values
+	if tableLog < minTablelog {
+		tableLog = minTablelog
+	}
+	if tableLog > tableLogMax {
+		tableLog = tableLogMax
+	}
+	s.actualTableLog = tableLog
+}
+
+type cTableEntry struct {
+	val   uint16
+	nBits uint8
+	// We have 8 bits extra
+}
+
+const huffNodesMask = huffNodesLen - 1
+
+func (s *Scratch) buildCTable() error {
+	s.huffSort()
+	if cap(s.cTable) < maxSymbolValue+1 {
+		s.cTable = make([]cTableEntry, s.symbolLen, maxSymbolValue+1)
+	} else {
+		s.cTable = s.cTable[:s.symbolLen]
+		for i := range s.cTable {
+			s.cTable[i] = cTableEntry{}
+		}
+	}
+
+	var startNode = int16(s.symbolLen)
+	nonNullRank := s.symbolLen - 1
+
+	nodeNb := int16(startNode)
+	huffNode := s.nodes[1 : huffNodesLen+1]
+
+	// This overlays the slice above, but allows "-1" index lookups.
+	// Different from reference implementation.
+	huffNode0 := s.nodes[0 : huffNodesLen+1]
+
+	for huffNode[nonNullRank].count == 0 {
+		nonNullRank--
+	}
+
+	lowS := int16(nonNullRank)
+	nodeRoot := nodeNb + lowS - 1
+	lowN := nodeNb
+	huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count
+	huffNode[lowS].parent, huffNode[lowS-1].parent = uint16(nodeNb), uint16(nodeNb)
+	nodeNb++
+	lowS -= 2
+	for n := nodeNb; n <= nodeRoot; n++ {
+		huffNode[n].count = 1 << 30
+	}
+	// fake entry, strong barrier
+	huffNode0[0].count = 1 << 31
+
+	// create parents
+	for nodeNb <= nodeRoot {
+		var n1, n2 int16
+		if huffNode0[lowS+1].count < huffNode0[lowN+1].count {
+			n1 = lowS
+			lowS--
+		} else {
+			n1 = lowN
+			lowN++
+		}
+		if huffNode0[lowS+1].count < huffNode0[lowN+1].count {
+			n2 = lowS
+			lowS--
+		} else {
+			n2 = lowN
+			lowN++
+		}
+
+		huffNode[nodeNb].count = huffNode0[n1+1].count + huffNode0[n2+1].count
+		huffNode0[n1+1].parent, huffNode0[n2+1].parent = uint16(nodeNb), uint16(nodeNb)
+		nodeNb++
+	}
+
+	// distribute weights (unlimited tree height)
+	huffNode[nodeRoot].nbBits = 0
+	for n := nodeRoot - 1; n >= startNode; n-- {
+		huffNode[n].nbBits = huffNode[huffNode[n].parent].nbBits + 1
+	}
+	for n := uint16(0); n <= nonNullRank; n++ {
+		huffNode[n].nbBits = huffNode[huffNode[n].parent].nbBits + 1
+	}
+	s.actualTableLog = s.setMaxHeight(int(nonNullRank))
+	maxNbBits := s.actualTableLog
+
+	// fill result into tree (val, nbBits)
+	if maxNbBits > tableLogMax {
+		return fmt.Errorf("internal error: maxNbBits (%d) > tableLogMax (%d)", maxNbBits, tableLogMax)
+	}
+	var nbPerRank [tableLogMax + 1]uint16
+	var valPerRank [tableLogMax + 1]uint16
+	for _, v := range huffNode[:nonNullRank+1] {
+		nbPerRank[v.nbBits]++
+	}
+	// determine stating value per rank
+	{
+		min := uint16(0)
+		for n := maxNbBits; n > 0; n-- {
+			// get starting value within each rank
+			valPerRank[n] = min
+			min += nbPerRank[n]
+			min >>= 1
+		}
+	}
+
+	// push nbBits per symbol, symbol order
+	// TODO: changed `s.symbolLen` -> `nonNullRank+1` (micro-opt)
+	for _, v := range huffNode[:nonNullRank+1] {
+		s.cTable[v.symbol].nBits = v.nbBits
+	}
+
+	// assign value within rank, symbol order
+	for n, val := range s.cTable[:s.symbolLen] {
+		v := valPerRank[val.nBits]
+		s.cTable[n].val = v
+		valPerRank[val.nBits] = v + 1
+	}
+
+	return nil
+}
+
+// huffSort will sort symbols, decreasing order.
+func (s *Scratch) huffSort() {
+	type rankPos struct {
+		base    uint32
+		current uint32
+	}
+
+	// Clear nodes
+	nodes := s.nodes[:huffNodesLen+1]
+	s.nodes = nodes
+	nodes = nodes[1 : huffNodesLen+1]
+
+	// Sort into buckets based on length of symbol count.
+	var rank [32]rankPos
+	for _, v := range s.count[:s.symbolLen] {
+		r := highBit32(v+1) & 31
+		rank[r].base++
+	}
+	for n := 30; n > 0; n-- {
+		rank[n-1].base += rank[n].base
+	}
+	for n := range rank[:] {
+		rank[n].current = rank[n].base
+	}
+	for n, c := range s.count[:s.symbolLen] {
+		r := (highBit32(c+1) + 1) & 31
+		pos := rank[r].current
+		rank[r].current++
+		prev := nodes[(pos-1)&huffNodesMask]
+		for pos > rank[r].base && c > prev.count {
+			nodes[pos&huffNodesMask] = prev
+			pos--
+			prev = nodes[(pos-1)&huffNodesMask]
+		}
+		nodes[pos&huffNodesMask] = nodeElt{count: c, symbol: byte(n)}
+	}
+	return
+}
+
+func (s *Scratch) setMaxHeight(lastNonNull int) uint8 {
+	maxNbBits := s.TableLog
+	huffNode := s.nodes[1 : huffNodesLen+1]
+	//huffNode = huffNode[: huffNodesLen]
+
+	largestBits := huffNode[lastNonNull].nbBits
+
+	// early exit : no elt > maxNbBits
+	if largestBits <= maxNbBits {
+		return largestBits
+	}
+	totalCost := int(0)
+	baseCost := int(1) << (largestBits - maxNbBits)
+	n := uint32(lastNonNull)
+
+	for huffNode[n].nbBits > maxNbBits {
+		totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits))
+		huffNode[n].nbBits = maxNbBits
+		n--
+	}
+	// n stops at huffNode[n].nbBits <= maxNbBits
+
+	for huffNode[n].nbBits == maxNbBits {
+		n--
+	}
+	// n end at index of smallest symbol using < maxNbBits
+
+	// renorm totalCost
+	totalCost >>= largestBits - maxNbBits /* note : totalCost is necessarily a multiple of baseCost */
+
+	// repay normalized cost
+	{
+		const noSymbol = 0xF0F0F0F0
+		var rankLast [tableLogMax + 2]uint32
+
+		for i := range rankLast[:] {
+			rankLast[i] = noSymbol
+		}
+
+		// Get pos of last (smallest) symbol per rank
+		{
+			currentNbBits := uint8(maxNbBits)
+			for pos := int(n); pos >= 0; pos-- {
+				if huffNode[pos].nbBits >= currentNbBits {
+					continue
+				}
+				currentNbBits = huffNode[pos].nbBits // < maxNbBits
+				rankLast[maxNbBits-currentNbBits] = uint32(pos)
+			}
+		}
+
+		for totalCost > 0 {
+			nBitsToDecrease := uint8(highBit32(uint32(totalCost))) + 1
+
+			for ; nBitsToDecrease > 1; nBitsToDecrease-- {
+				highPos := rankLast[nBitsToDecrease]
+				lowPos := rankLast[nBitsToDecrease-1]
+				if highPos == noSymbol {
+					continue
+				}
+				if lowPos == noSymbol {
+					break
+				}
+				highTotal := huffNode[highPos].count
+				lowTotal := 2 * huffNode[lowPos].count
+				if highTotal <= lowTotal {
+					break
+				}
+			}
+			// only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !)
+			// HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary
+			// FIXME: try to remove
+			for (nBitsToDecrease <= tableLogMax) && (rankLast[nBitsToDecrease] == noSymbol) {
+				nBitsToDecrease++
+			}
+			totalCost -= 1 << (nBitsToDecrease - 1)
+			if rankLast[nBitsToDecrease-1] == noSymbol {
+				// this rank is no longer empty
+				rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease]
+			}
+			huffNode[rankLast[nBitsToDecrease]].nbBits++
+			if rankLast[nBitsToDecrease] == 0 {
+				/* special case, reached largest symbol */
+				rankLast[nBitsToDecrease] = noSymbol
+			} else {
+				rankLast[nBitsToDecrease]--
+				if huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease {
+					rankLast[nBitsToDecrease] = noSymbol /* this rank is now empty */
+				}
+			}
+		}
+
+		for totalCost < 0 { /* Sometimes, cost correction overshoot */
+			if rankLast[1] == noSymbol { /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */
+				for huffNode[n].nbBits == maxNbBits {
+					n--
+				}
+				huffNode[n+1].nbBits--
+				rankLast[1] = n + 1
+				totalCost++
+				continue
+			}
+			huffNode[rankLast[1]+1].nbBits--
+			rankLast[1]++
+			totalCost++
+		}
+	}
+	return maxNbBits
+}
+
+type nodeElt struct {
+	count  uint32
+	parent uint16
+	symbol byte
+	nbBits uint8
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress.go b/vendor/github.com/klauspost/compress/huff0/decompress.go
new file mode 100644
index 0000000..43b4815
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress.go
@@ -0,0 +1,416 @@
+package huff0
+
+import (
+	"errors"
+	"fmt"
+	"io"
+
+	"github.com/klauspost/compress/fse"
+)
+
+type dTable struct {
+	single []dEntrySingle
+	double []dEntryDouble
+}
+
+// single-symbols decoding
+type dEntrySingle struct {
+	byte  uint8
+	nBits uint8
+}
+
+// double-symbols decoding
+type dEntryDouble struct {
+	seq   uint16
+	nBits uint8
+	len   uint8
+}
+
+// ReadTable will read a table from the input.
+// The size of the input may be larger than the table definition.
+// Any content remaining after the table definition will be returned.
+// If no Scratch is provided a new one is allocated.
+// The returned Scratch can be used for decoding input using this table.
+func ReadTable(in []byte, s *Scratch) (s2 *Scratch, remain []byte, err error) {
+	s, err = s.prepare(in)
+	if err != nil {
+		return s, nil, err
+	}
+	if len(in) <= 1 {
+		return s, nil, errors.New("input too small for table")
+	}
+	iSize := in[0]
+	in = in[1:]
+	if iSize >= 128 {
+		// Uncompressed
+		oSize := iSize - 127
+		iSize = (oSize + 1) / 2
+		if int(iSize) > len(in) {
+			return s, nil, errors.New("input too small for table")
+		}
+		for n := uint8(0); n < oSize; n += 2 {
+			v := in[n/2]
+			s.huffWeight[n] = v >> 4
+			s.huffWeight[n+1] = v & 15
+		}
+		s.symbolLen = uint16(oSize)
+		in = in[iSize:]
+	} else {
+		if len(in) <= int(iSize) {
+			return s, nil, errors.New("input too small for table")
+		}
+		// FSE compressed weights
+		s.fse.DecompressLimit = 255
+		hw := s.huffWeight[:]
+		s.fse.Out = hw
+		b, err := fse.Decompress(in[:iSize], s.fse)
+		s.fse.Out = nil
+		if err != nil {
+			return s, nil, err
+		}
+		if len(b) > 255 {
+			return s, nil, errors.New("corrupt input: output table too large")
+		}
+		s.symbolLen = uint16(len(b))
+		in = in[iSize:]
+	}
+
+	// collect weight stats
+	var rankStats [tableLogMax + 1]uint32
+	weightTotal := uint32(0)
+	for _, v := range s.huffWeight[:s.symbolLen] {
+		if v > tableLogMax {
+			return s, nil, errors.New("corrupt input: weight too large")
+		}
+		rankStats[v]++
+		weightTotal += (1 << (v & 15)) >> 1
+	}
+	if weightTotal == 0 {
+		return s, nil, errors.New("corrupt input: weights zero")
+	}
+
+	// get last non-null symbol weight (implied, total must be 2^n)
+	{
+		tableLog := highBit32(weightTotal) + 1
+		if tableLog > tableLogMax {
+			return s, nil, errors.New("corrupt input: tableLog too big")
+		}
+		s.actualTableLog = uint8(tableLog)
+		// determine last weight
+		{
+			total := uint32(1) << tableLog
+			rest := total - weightTotal
+			verif := uint32(1) << highBit32(rest)
+			lastWeight := highBit32(rest) + 1
+			if verif != rest {
+				// last value must be a clean power of 2
+				return s, nil, errors.New("corrupt input: last value not power of two")
+			}
+			s.huffWeight[s.symbolLen] = uint8(lastWeight)
+			s.symbolLen++
+			rankStats[lastWeight]++
+		}
+	}
+
+	if (rankStats[1] < 2) || (rankStats[1]&1 != 0) {
+		// by construction : at least 2 elts of rank 1, must be even
+		return s, nil, errors.New("corrupt input: min elt size, even check failed ")
+	}
+
+	// TODO: Choose between single/double symbol decoding
+
+	// Calculate starting value for each rank
+	{
+		var nextRankStart uint32
+		for n := uint8(1); n < s.actualTableLog+1; n++ {
+			current := nextRankStart
+			nextRankStart += rankStats[n] << (n - 1)
+			rankStats[n] = current
+		}
+	}
+
+	// fill DTable (always full size)
+	tSize := 1 << tableLogMax
+	if len(s.dt.single) != tSize {
+		s.dt.single = make([]dEntrySingle, tSize)
+	}
+
+	for n, w := range s.huffWeight[:s.symbolLen] {
+		length := (uint32(1) << w) >> 1
+		d := dEntrySingle{
+			byte:  uint8(n),
+			nBits: s.actualTableLog + 1 - w,
+		}
+		for u := rankStats[w]; u < rankStats[w]+length; u++ {
+			s.dt.single[u] = d
+		}
+		rankStats[w] += length
+	}
+	return s, in, nil
+}
+
+// Decompress1X will decompress a 1X encoded stream.
+// The length of the supplied input must match the end of a block exactly.
+// Before this is called, the table must be initialized with ReadTable unless
+// the encoder re-used the table.
+func (s *Scratch) Decompress1X(in []byte) (out []byte, err error) {
+	if len(s.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	var br bitReader
+	err = br.init(in)
+	if err != nil {
+		return nil, err
+	}
+	s.Out = s.Out[:0]
+
+	decode := func() byte {
+		val := br.peekBitsFast(s.actualTableLog) /* note : actualTableLog >= 1 */
+		v := s.dt.single[val]
+		br.bitsRead += v.nBits
+		return v.byte
+	}
+	hasDec := func(v dEntrySingle) byte {
+		br.bitsRead += v.nBits
+		return v.byte
+	}
+
+	// Avoid bounds check by always having full sized table.
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	dt := s.dt.single[:tlSize]
+
+	// Use temp table to avoid bound checks/append penalty.
+	var tmp = s.huffWeight[:256]
+	var off uint8
+
+	for br.off >= 8 {
+		br.fillFast()
+		tmp[off+0] = hasDec(dt[br.peekBitsFast(s.actualTableLog)&tlMask])
+		tmp[off+1] = hasDec(dt[br.peekBitsFast(s.actualTableLog)&tlMask])
+		br.fillFast()
+		tmp[off+2] = hasDec(dt[br.peekBitsFast(s.actualTableLog)&tlMask])
+		tmp[off+3] = hasDec(dt[br.peekBitsFast(s.actualTableLog)&tlMask])
+		off += 4
+		if off == 0 {
+			if len(s.Out)+256 > s.MaxDecodedSize {
+				br.close()
+				return nil, ErrMaxDecodedSizeExceeded
+			}
+			s.Out = append(s.Out, tmp...)
+		}
+	}
+
+	if len(s.Out)+int(off) > s.MaxDecodedSize {
+		br.close()
+		return nil, ErrMaxDecodedSizeExceeded
+	}
+	s.Out = append(s.Out, tmp[:off]...)
+
+	for !br.finished() {
+		br.fill()
+		if len(s.Out) >= s.MaxDecodedSize {
+			br.close()
+			return nil, ErrMaxDecodedSizeExceeded
+		}
+		s.Out = append(s.Out, decode())
+	}
+	return s.Out, br.close()
+}
+
+// Decompress4X will decompress a 4X encoded stream.
+// Before this is called, the table must be initialized with ReadTable unless
+// the encoder re-used the table.
+// The length of the supplied input must match the end of a block exactly.
+// The destination size of the uncompressed data must be known and provided.
+func (s *Scratch) Decompress4X(in []byte, dstSize int) (out []byte, err error) {
+	if len(s.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	if len(in) < 6+(4*1) {
+		return nil, errors.New("input too small")
+	}
+	if dstSize > s.MaxDecodedSize {
+		return nil, ErrMaxDecodedSizeExceeded
+	}
+	// TODO: We do not detect when we overrun a buffer, except if the last one does.
+
+	var br [4]bitReader
+	start := 6
+	for i := 0; i < 3; i++ {
+		length := int(in[i*2]) | (int(in[i*2+1]) << 8)
+		if start+length >= len(in) {
+			return nil, errors.New("truncated input (or invalid offset)")
+		}
+		err = br[i].init(in[start : start+length])
+		if err != nil {
+			return nil, err
+		}
+		start += length
+	}
+	err = br[3].init(in[start:])
+	if err != nil {
+		return nil, err
+	}
+
+	// Prepare output
+	if cap(s.Out) < dstSize {
+		s.Out = make([]byte, 0, dstSize)
+	}
+	s.Out = s.Out[:dstSize]
+	// destination, offset to match first output
+	dstOut := s.Out
+	dstEvery := (dstSize + 3) / 4
+
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	single := s.dt.single[:tlSize]
+
+	decode := func(br *bitReader) byte {
+		val := br.peekBitsFast(s.actualTableLog) /* note : actualTableLog >= 1 */
+		v := single[val&tlMask]
+		br.bitsRead += v.nBits
+		return v.byte
+	}
+
+	// Use temp table to avoid bound checks/append penalty.
+	var tmp = s.huffWeight[:256]
+	var off uint8
+
+	// Decode 2 values from each decoder/loop.
+	const bufoff = 256 / 4
+bigloop:
+	for {
+		for i := range br {
+			if br[i].off < 4 {
+				break bigloop
+			}
+			br[i].fillFast()
+		}
+		tmp[off] = decode(&br[0])
+		tmp[off+bufoff] = decode(&br[1])
+		tmp[off+bufoff*2] = decode(&br[2])
+		tmp[off+bufoff*3] = decode(&br[3])
+		tmp[off+1] = decode(&br[0])
+		tmp[off+1+bufoff] = decode(&br[1])
+		tmp[off+1+bufoff*2] = decode(&br[2])
+		tmp[off+1+bufoff*3] = decode(&br[3])
+		off += 2
+		if off == bufoff {
+			if bufoff > dstEvery {
+				return nil, errors.New("corruption detected: stream overrun 1")
+			}
+			copy(dstOut, tmp[:bufoff])
+			copy(dstOut[dstEvery:], tmp[bufoff:bufoff*2])
+			copy(dstOut[dstEvery*2:], tmp[bufoff*2:bufoff*3])
+			copy(dstOut[dstEvery*3:], tmp[bufoff*3:bufoff*4])
+			off = 0
+			dstOut = dstOut[bufoff:]
+			// There must at least be 3 buffers left.
+			if len(dstOut) < dstEvery*3 {
+				return nil, errors.New("corruption detected: stream overrun 2")
+			}
+		}
+	}
+	if off > 0 {
+		ioff := int(off)
+		if len(dstOut) < dstEvery*3+ioff {
+			return nil, errors.New("corruption detected: stream overrun 3")
+		}
+		copy(dstOut, tmp[:off])
+		copy(dstOut[dstEvery:dstEvery+ioff], tmp[bufoff:bufoff*2])
+		copy(dstOut[dstEvery*2:dstEvery*2+ioff], tmp[bufoff*2:bufoff*3])
+		copy(dstOut[dstEvery*3:dstEvery*3+ioff], tmp[bufoff*3:bufoff*4])
+		dstOut = dstOut[off:]
+	}
+
+	for i := range br {
+		offset := dstEvery * i
+		br := &br[i]
+		for !br.finished() {
+			br.fill()
+			if offset >= len(dstOut) {
+				return nil, errors.New("corruption detected: stream overrun 4")
+			}
+			dstOut[offset] = decode(br)
+			offset++
+		}
+		err = br.close()
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	return s.Out, nil
+}
+
+// matches will compare a decoding table to a coding table.
+// Errors are written to the writer.
+// Nothing will be written if table is ok.
+func (s *Scratch) matches(ct cTable, w io.Writer) {
+	if s == nil || len(s.dt.single) == 0 {
+		return
+	}
+	dt := s.dt.single[:1<<s.actualTableLog]
+	tablelog := s.actualTableLog
+	ok := 0
+	broken := 0
+	for sym, enc := range ct {
+		errs := 0
+		broken++
+		if enc.nBits == 0 {
+			for _, dec := range dt {
+				if dec.byte == byte(sym) {
+					fmt.Fprintf(w, "symbol %x has decoder, but no encoder\n", sym)
+					errs++
+					break
+				}
+			}
+			if errs == 0 {
+				broken--
+			}
+			continue
+		}
+		// Unused bits in input
+		ub := tablelog - enc.nBits
+		top := enc.val << ub
+		// decoder looks at top bits.
+		dec := dt[top]
+		if dec.nBits != enc.nBits {
+			fmt.Fprintf(w, "symbol 0x%x bit size mismatch (enc: %d, dec:%d).\n", sym, enc.nBits, dec.nBits)
+			errs++
+		}
+		if dec.byte != uint8(sym) {
+			fmt.Fprintf(w, "symbol 0x%x decoder output mismatch (enc: %d, dec:%d).\n", sym, sym, dec.byte)
+			errs++
+		}
+		if errs > 0 {
+			fmt.Fprintf(w, "%d errros in base, stopping\n", errs)
+			continue
+		}
+		// Ensure that all combinations are covered.
+		for i := uint16(0); i < (1 << ub); i++ {
+			vval := top | i
+			dec := dt[vval]
+			if dec.nBits != enc.nBits {
+				fmt.Fprintf(w, "symbol 0x%x bit size mismatch (enc: %d, dec:%d).\n", vval, enc.nBits, dec.nBits)
+				errs++
+			}
+			if dec.byte != uint8(sym) {
+				fmt.Fprintf(w, "symbol 0x%x decoder output mismatch (enc: %d, dec:%d).\n", vval, sym, dec.byte)
+				errs++
+			}
+			if errs > 20 {
+				fmt.Fprintf(w, "%d errros, stopping\n", errs)
+				break
+			}
+		}
+		if errs == 0 {
+			ok++
+			broken--
+		}
+	}
+	if broken > 0 {
+		fmt.Fprintf(w, "%d broken, %d ok\n", broken, ok)
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/huff0.go b/vendor/github.com/klauspost/compress/huff0/huff0.go
new file mode 100644
index 0000000..6f823f9
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/huff0.go
@@ -0,0 +1,252 @@
+// Package huff0 provides fast huffman encoding as used in zstd.
+//
+// See README.md at https://github.com/klauspost/compress/tree/master/huff0 for details.
+package huff0
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"math/bits"
+
+	"github.com/klauspost/compress/fse"
+)
+
+const (
+	maxSymbolValue = 255
+
+	// zstandard limits tablelog to 11, see:
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#huffman-tree-description
+	tableLogMax     = 11
+	tableLogDefault = 11
+	minTablelog     = 5
+	huffNodesLen    = 512
+
+	// BlockSizeMax is maximum input size for a single block uncompressed.
+	BlockSizeMax = 1<<18 - 1
+)
+
+var (
+	// ErrIncompressible is returned when input is judged to be too hard to compress.
+	ErrIncompressible = errors.New("input is not compressible")
+
+	// ErrUseRLE is returned from the compressor when the input is a single byte value repeated.
+	ErrUseRLE = errors.New("input is single value repeated")
+
+	// ErrTooBig is return if input is too large for a single block.
+	ErrTooBig = errors.New("input too big")
+
+	// ErrMaxDecodedSizeExceeded is return if input is too large for a single block.
+	ErrMaxDecodedSizeExceeded = errors.New("maximum output size exceeded")
+)
+
+type ReusePolicy uint8
+
+const (
+	// ReusePolicyAllow will allow reuse if it produces smaller output.
+	ReusePolicyAllow ReusePolicy = iota
+
+	// ReusePolicyPrefer will re-use aggressively if possible.
+	// This will not check if a new table will produce smaller output,
+	// except if the current table is impossible to use or
+	// compressed output is bigger than input.
+	ReusePolicyPrefer
+
+	// ReusePolicyNone will disable re-use of tables.
+	// This is slightly faster than ReusePolicyAllow but may produce larger output.
+	ReusePolicyNone
+)
+
+type Scratch struct {
+	count [maxSymbolValue + 1]uint32
+
+	// Per block parameters.
+	// These can be used to override compression parameters of the block.
+	// Do not touch, unless you know what you are doing.
+
+	// Out is output buffer.
+	// If the scratch is re-used before the caller is done processing the output,
+	// set this field to nil.
+	// Otherwise the output buffer will be re-used for next Compression/Decompression step
+	// and allocation will be avoided.
+	Out []byte
+
+	// OutTable will contain the table data only, if a new table has been generated.
+	// Slice of the returned data.
+	OutTable []byte
+
+	// OutData will contain the compressed data.
+	// Slice of the returned data.
+	OutData []byte
+
+	// MaxSymbolValue will override the maximum symbol value of the next block.
+	MaxSymbolValue uint8
+
+	// TableLog will attempt to override the tablelog for the next block.
+	// Must be <= 11.
+	TableLog uint8
+
+	// Reuse will specify the reuse policy
+	Reuse ReusePolicy
+
+	// MaxDecodedSize will set the maximum allowed output size.
+	// This value will automatically be set to BlockSizeMax if not set.
+	// Decoders will return ErrMaxDecodedSizeExceeded is this limit is exceeded.
+	MaxDecodedSize int
+
+	br             byteReader
+	symbolLen      uint16 // Length of active part of the symbol table.
+	maxCount       int    // count of the most probable symbol
+	clearCount     bool   // clear count
+	actualTableLog uint8  // Selected tablelog.
+	prevTable      cTable // Table used for previous compression.
+	cTable         cTable // compression table
+	dt             dTable // decompression table
+	nodes          []nodeElt
+	tmpOut         [4][]byte
+	fse            *fse.Scratch
+	huffWeight     [maxSymbolValue + 1]byte
+}
+
+func (s *Scratch) prepare(in []byte) (*Scratch, error) {
+	if len(in) > BlockSizeMax {
+		return nil, ErrTooBig
+	}
+	if s == nil {
+		s = &Scratch{}
+	}
+	if s.MaxSymbolValue == 0 {
+		s.MaxSymbolValue = maxSymbolValue
+	}
+	if s.TableLog == 0 {
+		s.TableLog = tableLogDefault
+	}
+	if s.TableLog > tableLogMax {
+		return nil, fmt.Errorf("tableLog (%d) > maxTableLog (%d)", s.TableLog, tableLogMax)
+	}
+	if s.MaxDecodedSize <= 0 || s.MaxDecodedSize > BlockSizeMax {
+		s.MaxDecodedSize = BlockSizeMax
+	}
+	if s.clearCount && s.maxCount == 0 {
+		for i := range s.count {
+			s.count[i] = 0
+		}
+		s.clearCount = false
+	}
+	if cap(s.Out) == 0 {
+		s.Out = make([]byte, 0, len(in))
+	}
+	s.Out = s.Out[:0]
+
+	s.OutTable = nil
+	s.OutData = nil
+	if cap(s.nodes) < huffNodesLen+1 {
+		s.nodes = make([]nodeElt, 0, huffNodesLen+1)
+	}
+	s.nodes = s.nodes[:0]
+	if s.fse == nil {
+		s.fse = &fse.Scratch{}
+	}
+	s.br.init(in)
+
+	return s, nil
+}
+
+type cTable []cTableEntry
+
+func (c cTable) write(s *Scratch) error {
+	var (
+		// precomputed conversion table
+		bitsToWeight [tableLogMax + 1]byte
+		huffLog      = s.actualTableLog
+		// last weight is not saved.
+		maxSymbolValue = uint8(s.symbolLen - 1)
+		huffWeight     = s.huffWeight[:256]
+	)
+	const (
+		maxFSETableLog = 6
+	)
+	// convert to weight
+	bitsToWeight[0] = 0
+	for n := uint8(1); n < huffLog+1; n++ {
+		bitsToWeight[n] = huffLog + 1 - n
+	}
+
+	// Acquire histogram for FSE.
+	hist := s.fse.Histogram()
+	hist = hist[:256]
+	for i := range hist[:16] {
+		hist[i] = 0
+	}
+	for n := uint8(0); n < maxSymbolValue; n++ {
+		v := bitsToWeight[c[n].nBits] & 15
+		huffWeight[n] = v
+		hist[v]++
+	}
+
+	// FSE compress if feasible.
+	if maxSymbolValue >= 2 {
+		huffMaxCnt := uint32(0)
+		huffMax := uint8(0)
+		for i, v := range hist[:16] {
+			if v == 0 {
+				continue
+			}
+			huffMax = byte(i)
+			if v > huffMaxCnt {
+				huffMaxCnt = v
+			}
+		}
+		s.fse.HistogramFinished(huffMax, int(huffMaxCnt))
+		s.fse.TableLog = maxFSETableLog
+		b, err := fse.Compress(huffWeight[:maxSymbolValue], s.fse)
+		if err == nil && len(b) < int(s.symbolLen>>1) {
+			s.Out = append(s.Out, uint8(len(b)))
+			s.Out = append(s.Out, b...)
+			return nil
+		}
+		// Unable to compress (RLE/uncompressible)
+	}
+	// write raw values as 4-bits (max : 15)
+	if maxSymbolValue > (256 - 128) {
+		// should not happen : likely means source cannot be compressed
+		return ErrIncompressible
+	}
+	op := s.Out
+	// special case, pack weights 4 bits/weight.
+	op = append(op, 128|(maxSymbolValue-1))
+	// be sure it doesn't cause msan issue in final combination
+	huffWeight[maxSymbolValue] = 0
+	for n := uint16(0); n < uint16(maxSymbolValue); n += 2 {
+		op = append(op, (huffWeight[n]<<4)|huffWeight[n+1])
+	}
+	s.Out = op
+	return nil
+}
+
+// estimateSize returns the estimated size in bytes of the input represented in the
+// histogram supplied.
+func (c cTable) estimateSize(hist []uint32) int {
+	nbBits := uint32(7)
+	for i, v := range c[:len(hist)] {
+		nbBits += uint32(v.nBits) * hist[i]
+	}
+	return int(nbBits >> 3)
+}
+
+// minSize returns the minimum possible size considering the shannon limit.
+func (s *Scratch) minSize(total int) int {
+	nbBits := float64(7)
+	fTotal := float64(total)
+	for _, v := range s.count[:s.symbolLen] {
+		n := float64(v)
+		if n > 0 {
+			nbBits += math.Log2(fTotal/n) * n
+		}
+	}
+	return int(nbBits) >> 3
+}
+
+func highBit32(val uint32) (n uint32) {
+	return uint32(bits.Len32(val) - 1)
+}